Merge pull request #485 from modelscope/usp

support Unified Sequence Parallel

diffsynth/distributed/xdit_context_parallel.py

@@ -0,0 +1,127 @@
+import torch
+from typing import Optional
+from einops import rearrange
+from xfuser.core.distributed import (get_sequence_parallel_rank,
+                                     get_sequence_parallel_world_size,
+                                     get_sp_group)
+from xfuser.core.long_ctx_attention import xFuserLongContextAttention
+
+def sinusoidal_embedding_1d(dim, position):
+    sinusoid = torch.outer(position.type(torch.float64), torch.pow(
+        10000, -torch.arange(dim//2, dtype=torch.float64, device=position.device).div(dim//2)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x.to(position.dtype)
+
+def pad_freqs(original_tensor, target_len):
+    seq_len, s1, s2 = original_tensor.shape
+    pad_size = target_len - seq_len
+    padding_tensor = torch.ones(
+        pad_size,
+        s1,
+        s2,
+        dtype=original_tensor.dtype,
+        device=original_tensor.device)
+    padded_tensor = torch.cat([original_tensor, padding_tensor], dim=0)
+    return padded_tensor
+    
+def rope_apply(x, freqs, num_heads):
+    x = rearrange(x, "b s (n d) -> b s n d", n=num_heads)
+    s_per_rank = x.shape[1]
+
+    x_out = torch.view_as_complex(x.to(torch.float64).reshape(
+        x.shape[0], x.shape[1], x.shape[2], -1, 2))
+
+    sp_size = get_sequence_parallel_world_size()
+    sp_rank = get_sequence_parallel_rank()
+    freqs = pad_freqs(freqs, s_per_rank * sp_size)
+    freqs_rank = freqs[(sp_rank * s_per_rank):((sp_rank + 1) * s_per_rank), :, :]
+
+    x_out = torch.view_as_real(x_out * freqs_rank).flatten(2)
+    return x_out.to(x.dtype)
+
+def usp_dit_forward(self,
+            x: torch.Tensor,
+            timestep: torch.Tensor,
+            context: torch.Tensor,
+            clip_feature: Optional[torch.Tensor] = None,
+            y: Optional[torch.Tensor] = None,
+            use_gradient_checkpointing: bool = False,
+            use_gradient_checkpointing_offload: bool = False,
+            **kwargs,
+            ):
+    t = self.time_embedding(
+        sinusoidal_embedding_1d(self.freq_dim, timestep))
+    t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
+    context = self.text_embedding(context)
+    
+    if self.has_image_input:
+        x = torch.cat([x, y], dim=1)  # (b, c_x + c_y, f, h, w)
+        clip_embdding = self.img_emb(clip_feature)
+        context = torch.cat([clip_embdding, context], dim=1)
+    
+    x, (f, h, w) = self.patchify(x)
+    
+    freqs = torch.cat([
+        self.freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+        self.freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+        self.freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+    ], dim=-1).reshape(f * h * w, 1, -1).to(x.device)
+    
+    def create_custom_forward(module):
+        def custom_forward(*inputs):
+            return module(*inputs)
+        return custom_forward
+
+    # Context Parallel
+    x = torch.chunk(
+        x, get_sequence_parallel_world_size(),
+        dim=1)[get_sequence_parallel_rank()]
+
+    for block in self.blocks:
+        if self.training and use_gradient_checkpointing:
+            if use_gradient_checkpointing_offload:
+                with torch.autograd.graph.save_on_cpu():
+                    x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            else:
+                x = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    x, context, t_mod, freqs,
+                    use_reentrant=False,
+                )
+        else:
+            x = block(x, context, t_mod, freqs)
+
+    x = self.head(x, t)
+
+    # Context Parallel
+    x = get_sp_group().all_gather(x, dim=1)
+
+    # unpatchify
+    x = self.unpatchify(x, (f, h, w))
+    return x
+
+
+def usp_attn_forward(self, x, freqs):
+    q = self.norm_q(self.q(x))
+    k = self.norm_k(self.k(x))
+    v = self.v(x)
+
+    q = rope_apply(q, freqs, self.num_heads)
+    k = rope_apply(k, freqs, self.num_heads)
+    q = rearrange(q, "b s (n d) -> b s n d", n=self.num_heads)
+    k = rearrange(k, "b s (n d) -> b s n d", n=self.num_heads)
+    v = rearrange(v, "b s (n d) -> b s n d", n=self.num_heads)
+
+    x = xFuserLongContextAttention()(
+        None,
+        query=q,
+        key=k,
+        value=v,
+    )
+    x = x.flatten(2)
+
+    return self.o(x)

diffsynth/pipelines/wan_video.py

@@ -1,3 +1,4 @@
+import types
 from ..models import ModelManager
 from ..models.wan_video_dit import WanModel
 from ..models.wan_video_text_encoder import WanTextEncoder
@@ -30,9 +31,10 @@ class WanVideoPipeline(BasePipeline):
         self.image_encoder: WanImageEncoder = None
         self.dit: WanModel = None
         self.vae: WanVideoVAE = None
-        self.model_names = ['text_encoder', 'dit', 'vae']
+        self.model_names = ['text_encoder', 'dit', 'vae', 'image_encoder']
         self.height_division_factor = 16
         self.width_division_factor = 16
+        self.use_unified_sequence_parallel = False
 
 
     def enable_vram_management(self, num_persistent_param_in_dit=None):
@@ -135,11 +137,20 @@ class WanVideoPipeline(BasePipeline):
 
 
     @staticmethod
-    def from_model_manager(model_manager: ModelManager, torch_dtype=None, device=None):
+    def from_model_manager(model_manager: ModelManager, torch_dtype=None, device=None, use_usp=False):
         if device is None: device = model_manager.device
         if torch_dtype is None: torch_dtype = model_manager.torch_dtype
         pipe = WanVideoPipeline(device=device, torch_dtype=torch_dtype)
         pipe.fetch_models(model_manager)
+        if use_usp:
+            from xfuser.core.distributed import get_sequence_parallel_world_size
+            from ..distributed.xdit_context_parallel import usp_attn_forward, usp_dit_forward
+
+            for block in pipe.dit.blocks:
+                block.self_attn.forward = types.MethodType(usp_attn_forward, block.self_attn)
+            pipe.dit.forward = types.MethodType(usp_dit_forward, pipe.dit)
+            pipe.sp_size = get_sequence_parallel_world_size()
+            pipe.use_unified_sequence_parallel = True
         return pipe
     
     
@@ -191,6 +202,10 @@ class WanVideoPipeline(BasePipeline):
         return frames
     
     
+    def prepare_unified_sequence_parallel(self):
+        return {"use_unified_sequence_parallel": self.use_unified_sequence_parallel}
+
+
     @torch.no_grad()
     def __call__(
         self,
@@ -259,15 +274,18 @@ class WanVideoPipeline(BasePipeline):
         tea_cache_posi = {"tea_cache": TeaCache(num_inference_steps, rel_l1_thresh=tea_cache_l1_thresh, model_id=tea_cache_model_id) if tea_cache_l1_thresh is not None else None}
         tea_cache_nega = {"tea_cache": TeaCache(num_inference_steps, rel_l1_thresh=tea_cache_l1_thresh, model_id=tea_cache_model_id) if tea_cache_l1_thresh is not None else None}
         
+        # Unified Sequence Parallel
+        usp_kwargs = self.prepare_unified_sequence_parallel()
+
         # Denoise
         self.load_models_to_device(["dit"])
         for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
             timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
 
             # Inference
-            noise_pred_posi = model_fn_wan_video(self.dit, latents, timestep=timestep, **prompt_emb_posi, **image_emb, **extra_input, **tea_cache_posi)
+            noise_pred_posi = model_fn_wan_video(self.dit, latents, timestep=timestep, **prompt_emb_posi, **image_emb, **extra_input, **tea_cache_posi, **usp_kwargs)
             if cfg_scale != 1.0:
-                noise_pred_nega = model_fn_wan_video(self.dit, latents, timestep=timestep, **prompt_emb_nega, **image_emb, **extra_input, **tea_cache_nega)
+                noise_pred_nega = model_fn_wan_video(self.dit, latents, timestep=timestep, **prompt_emb_nega, **image_emb, **extra_input, **tea_cache_nega, **usp_kwargs)
                 noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
             else:
                 noise_pred = noise_pred_posi
@@ -346,8 +364,15 @@ def model_fn_wan_video(
     clip_feature: Optional[torch.Tensor] = None,
     y: Optional[torch.Tensor] = None,
     tea_cache: TeaCache = None,
+    use_unified_sequence_parallel: bool = False,
     **kwargs,
 ):
+    if use_unified_sequence_parallel:
+        import torch.distributed as dist
+        from xfuser.core.distributed import (get_sequence_parallel_rank,
+                                            get_sequence_parallel_world_size,
+                                            get_sp_group)
+    
     t = dit.time_embedding(sinusoidal_embedding_1d(dit.freq_dim, timestep))
     t_mod = dit.time_projection(t).unflatten(1, (6, dit.dim))
     context = dit.text_embedding(context)
@@ -375,11 +400,17 @@ def model_fn_wan_video(
         x = tea_cache.update(x)
     else:
         # blocks
+        if use_unified_sequence_parallel:
+            if dist.is_initialized() and dist.get_world_size() > 1:
+                x = torch.chunk(x, get_sequence_parallel_world_size(), dim=1)[get_sequence_parallel_rank()]
         for block in dit.blocks:
             x = block(x, context, t_mod, freqs)
         if tea_cache is not None:
             tea_cache.store(x)
 
     x = dit.head(x, t)
+    if use_unified_sequence_parallel:
+        if dist.is_initialized() and dist.get_world_size() > 1:
+            x = get_sp_group().all_gather(x, dim=1)
     x = dit.unpatchify(x, (f, h, w))
     return x

examples/wanvideo/README.md

@@ -49,6 +49,20 @@ We present a detailed table here. The model is tested on a single A100.
 
 https://github.com/user-attachments/assets/3908bc64-d451-485a-8b61-28f6d32dd92f
 
+### Parallel Inference
+
+1. Unified Sequence Parallel (USP)
+
+```bash
+pip install xfuser>=0.4.3
+```
+
+```bash
+torchrun --standalone --nproc_per_node=8 examples/wanvideo/wan_14b_text_to_video_usp.py
+```
+
+2. Tensor Parallel
+
 Tensor parallel module of Wan-Video-14B-T2V is still under development. An example script is provided in [`./wan_14b_text_to_video_tensor_parallel.py`](./wan_14b_text_to_video_tensor_parallel.py).
 
 ### Wan-Video-14B-I2V

examples/wanvideo/wan_14b_text_to_video_usp.py

@@ -0,0 +1,58 @@
+import torch
+from diffsynth import ModelManager, WanVideoPipeline, save_video, VideoData
+from modelscope import snapshot_download
+import torch.distributed as dist
+
+
+# Download models
+snapshot_download("Wan-AI/Wan2.1-T2V-14B", local_dir="models/Wan-AI/Wan2.1-T2V-14B")
+
+# Load models
+model_manager = ModelManager(device="cpu")
+model_manager.load_models(
+    [
+        [
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00001-of-00006.safetensors",
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00002-of-00006.safetensors",
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00003-of-00006.safetensors",
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00004-of-00006.safetensors",
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00005-of-00006.safetensors",
+            "models/Wan-AI/Wan2.1-T2V-14B/diffusion_pytorch_model-00006-of-00006.safetensors",
+        ],
+        "models/Wan-AI/Wan2.1-T2V-14B/models_t5_umt5-xxl-enc-bf16.pth",
+        "models/Wan-AI/Wan2.1-T2V-14B/Wan2.1_VAE.pth",
+    ],
+    torch_dtype=torch.float8_e4m3fn, # You can set `torch_dtype=torch.bfloat16` to disable FP8 quantization.
+)
+
+dist.init_process_group(
+    backend="nccl",
+    init_method="env://",
+)
+from xfuser.core.distributed import (initialize_model_parallel,
+                                     init_distributed_environment)
+init_distributed_environment(
+    rank=dist.get_rank(), world_size=dist.get_world_size())
+
+initialize_model_parallel(
+    sequence_parallel_degree=dist.get_world_size(),
+    ring_degree=1,
+    ulysses_degree=dist.get_world_size(),
+)
+torch.cuda.set_device(dist.get_rank())
+
+pipe = WanVideoPipeline.from_model_manager(model_manager, 
+                                           torch_dtype=torch.bfloat16, 
+                                           device=f"cuda:{dist.get_rank()}", 
+                                           use_usp=True if dist.get_world_size() > 1 else False)
+pipe.enable_vram_management(num_persistent_param_in_dit=None) # You can set `num_persistent_param_in_dit` to a small number to reduce VRAM required.
+
+# Text-to-video
+video = pipe(
+    prompt="一名宇航员身穿太空服，面朝镜头骑着一匹机械马在火星表面驰骋。红色的荒凉地表延伸至远方，点缀着巨大的陨石坑和奇特的岩石结构。机械马的步伐稳健，扬起微弱的尘埃，展现出未来科技与原始探索的完美结合。宇航员手持操控装置，目光坚定，仿佛正在开辟人类的新疆域。背景是深邃的宇宙和蔚蓝的地球，画面既科幻又充满希望，让人不禁畅想未来的星际生活。",
+    negative_prompt="色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走",
+    num_inference_steps=50,
+    seed=0, tiled=True
+)
+if dist.get_rank() == 0:
+    save_video(video, "video1.mp4", fps=25, quality=5)